期刊
NEURAL NETWORKS
卷 170, 期 -, 页码 167-175出版社
PERGAMON-ELSEVIER SCIENCE LTD
DOI: 10.1016/j.neunet.2023.11.015
关键词
CDnet2014 dataset; Feature pooling module; Foreground segmentation; Multi -scale feature encodings; Transposed convolutional neural network
This article introduces a foreground segmentation algorithm that solves the challenges of darkness, dynamic background information, and camera jitter by using a triplet CNN, a Transposed Convolutional Neural Network (TCNN), and a Features Pooling Module (FPM). The results show that the algorithm outperforms other state-of-the-art algorithms on the CDnet2014 datasets and enhances the foreground segmentation performance.
Foreground segmentation algorithm aims to precisely separate moving objects from the background in various environments. However, the interference from darkness, dynamic background information, and camera jitter makes it still challenging to build a decent detection network. To solve these issues, a triplet CNN and Transposed Convolutional Neural Network (TCNN) are created by attaching a Features Pooling Module (FPM). TCNN process reduces the amount of multi-scale inputs to the network by fusing features into the Foreground Segmentation Network (FgSegNet) based FPM, which extracts multi-scale features from images and builds a strong feature pooling. Additionally, the up-sampling network is added to the proposed technique, which is used to up-sample the abstract image representation, so that its spatial dimensions match with the input image. The large context and long-range dependencies among pixels are acquired by TCNN and segmentation mask, in multiple scales using triplet CNN, to enhance the foreground segmentation of FgSegNet. The results, clearly show that FgSegNet surpasses other state-of-the-art algorithms on the CDnet2014 datasets, with an average F-Measure of 0.9804, precision of 0.9801, PWC as (0.0461), and recall as (0.9896). Moreover, the FgSegNet with up-sampling achieves the F-measure of 0.9804 which is higher when compared to the FgSegNet without up-sampling.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据